Dominant preference decoupling and guided perturbed preference injection for cross-domain sequence recommendation

Cross-domain sequential recommendation jointly models cross- and intra-domain interaction sequences to extract relevant information to predict future interactions across domains. Nevertheless, current mainstream methods overlook the intra-domain dominant preference and the impact of perturbed preference on prediction outcomes. Hence, this paper proposes the Dominant Preference Decoupling and Guided Perturbed Preference Injection for Cross-Domain Sequence Recommendation (DP-CSR) model to address the aforementioned issues. The core idea is to preserve the intra-domain dominant preference while extracting perturbed preference information from cross-domain sequences to predict user interactions. Specifically, DP-CSR captures diverse intra-domain dominant preferences through multi-channel hypergraph learning and then integrates them using an attention mechanism. After that, it constructs serialized perturbed preference by jointly modeling intra and cross-domain sequences using sequence encoders. Furthermore, a gating mechanism dynamically injects critical cross-domain perturbed preference information into the intra-domain perturbed preference. This strategy enhances the model’s prediction adaptability by combining three preference types and avoiding information redundancy. Furthermore, a contrastive learning-based preference decoupling optimization objective enhances the preference decoupling and fine alignment of the cross-domain perturbed preferences with the intra-domain perturbed ones. Extensive experiments on six real-world benchmark datasets demonstrate remarkable and consistent improvements of the proposed DP-CSR over the state-of-the-art methods.